M3 Pro 18-core GPU (reported) LLM benchmark

On the M3 Pro 18-core GPU (reported), single-user chat decode speeds by model size (one result per model) are: ~7–9B models at 22.0 tok/s (mlx; community results only: middle of 3 community results, not measured by llm-speed, source run, community), ~14B models (too few results: 2 models, 0 by llm-speed, 2 community; no middle value) and ~32B models (too few results: 2 models, 0 by llm-speed, 2 community; no middle value). Short-chat workload, local runtimes, dense models only. Each model counts once (its standard run at default context), cold-load rows are left out, and a size class needs results from at least 3 models for a middle value. Quantization, backend and settings vary between submissions, so check each row in the tables below.